Oracle Database Discussions

Hello all;

So here is my world. We have central to our data monitoring system an oracle database running Oracle Standard One (please don't laugh... I understand it is comical) licensing.

This DB is about 1.7 TB of small record data.

One table in particular (the raw incoming data, 350gb, 8 billion rows, just in the table) is fed millions of rows each day in real time by two to three main "data collectors" or what have you. Data must be available in this table "as fast as possible" once it is received.

This table has 6 columns (one varchar usually empty, a few numerics including a source id, a timestamp and a create time).

The data is collect in chronological order (increasing timestamp) 90% of the time (though sometimes the timestamp may be very old and catch up to current). The other 10% of the time the data can be out of order according to the timestamp.

This table has two indexes, unique (sourceid, timestamp), and a non unique (create time). (FYI, this used to be an IOT until we had to add the second index on create time, at which point a secondary index on create time slowed the IOT to a crawl)

About 80% of this data is removed after it ages beyond 3 months; 20% is retained as "special" long term data (customer pays for longer raw source retention). The data is removed using delete statements. This table is never (99.99% of the time) updated. The indexes are not rebuilt... ever... as a rebuild is about a 20+ hour process, and without online rebuilds since we are standard one, this is just not possible.


Now what we are observing is that the inserts into this table

- Inserts are much slower based on a "wider" cardinality of the "sourceid" of the data being inserted. What I mean is that 10,000 inserts for 10,000 sourceid (regardless of timestamp) is MUCH, MUCH slower than 10,000 inserts for a single sourceid. This makes sense to me, as I understand it that oracle must inspect more branches of the index for uniqueness, and more different physical blocks will be used to store the new index data. There are about 2 million unique sourceId across our system.

- Over time, oracle is requesting more and more ram to satisfy these inserts in a timely matter. My understanding here is that oracle is attempting to hold the leafs of these indexes perpetually buffers. Our system does have a 99% cache hit rate. However, we are seeing oracle requiring roughly 10GB extra ram per quarter to 6 months; we're at about 50gb of ram just for oracle already.

- If I emulate our production load on a brand new, empty table / indexes, performance is easily 10x to 20x faster than what I see when I do the same tests with the large production copies of data.

We have the following assumption: Partitioning this table based on good logical grouping of sourceid, and then timestamp, will help reduce the work required by oracle to verify uniqueness of data, reducing the amount of data that must be cached by oracle, and allow us to handle our "older than 3 month" at a partition level, greatly reducing table and index fragmentation.

Based on our hardware, its going to be about a million dollar hit to upgrade to Enterprise (with partitioning), plus a couple hundred thousand a year in support. Currently I think we pay a whopping 5 grand a year in support, if that, total oracle costs. This is going to be a huge pill for our company to swallow.

What I am looking for guidance / help on, should we really expect partitioning to make a difference here? I want to get that 10x performance difference back we see between a fresh empty system, and our current production system. I also want to limit oracles 10gb / quarter growing need for more buffer cache (the cardinality of sourceid does NOT grow by that much per quarter... maybe 1000s per quarter, out of 2 million).

Also, please I'd appreciate it if there were no mocking comments about using standard one up to this point :) I know it is risky and insane and maybe more than a bit silly, but we make due with what we have. And all the credit in the world to oracle that their "entry" level system has been able to handle everything we've thrown at it so far! :)

Alright all, thank you very much for listening, and I look forward to hear the opinions of the experts.